Hotel Lobby AI face drift, swapped roles and occlusion: a small-change decision tree
This is a preparation decision tree, not a face detector or a proven repair. You choose the symptom and inspect the frames yourself. Keep the orange scene, two subjects and reference mapping explicit; do not add a long list of conflicting fixes.
Written and documentation checked
1. Keep the first bad frame as your comparison point
Check the input, the opening frame and the first point where the problem appears. If the problem exists in the input, repair that input before changing motion directions. If it begins later, compare the action or camera change immediately before it. Record the actual timestamp at the provider; DuoFrame does not read your video.
A fair small test keeps all other recorded settings the same, changes one variable and preserves the old version. It cannot prove that a random variation came from your change. Record failures as well as promising results, and do not generalize from a single clip.
2. Choose the smallest proposed change
| Symptom / when | Check first | Change only this next |
|---|---|---|
| Face drift / in input or first frame | Compare the actual selected reference with the visible subject. | Replace one unclear or obstructed reference; keep action/settings fixed. |
| Face drift / later | Find the first bad turn or fast action. | Reduce that one movement; keep the same reference mapping. |
| Roles swap / in input or first frame | Check source order and every left/right label. | Correct the mapping once; do not simultaneously replace photos. |
| Roles swap / later | Check whether performers cross or overlap. | Remove the crossing movement in the next brief. |
| Face blocked / in input or first frame | Inspect crop edges, microphone, hands and other subjects. | Reframe that input to reveal the face; do not claim object removal. |
| Face blocked / later | Locate the gesture that moves in front of the face. | Move that one gesture below the face in the written brief. |
3. Use observable directions instead of guarantees
An original, untested writing example: “The left subject makes one small head turn; the right subject stays on their side with hands below the face.” This describes a proposal you can inspect. “Perfect identity forever” does not tell you what to compare. Adapt element names to the actual provider rather than copying model-specific tags between tools.
For a finished starting frame, keep the appearance in the frame and make the motion instruction short. For source-video editing, confirm that the tool supports your input and describe preservation separately from replacement. A text-only draft cannot promise exact choreography, identity or lip sync.
4. Decide when to stop and collect better evidence
Stop repeating paid runs if you cannot tell whether the input is supported, a task is still processing, or the proposed change was actually applied. Keep the request ID and contact the provider through its support route. If every reference hides the face, a better authorized reference is more concrete than an invented identity score.
Save a named prompt version and export the brief with input order, one proposed change, model/version, settings and a place for the actual result. Our examples remain concepts until a permitted real run supplies input/output evidence. No face matching, biometric identification or automated confidence score is performed.
Sources and limits
Provider documentation supports the input distinctions. Practical checklists and example revisions here are original preparation advice. No real render, user study, success rate or cost result is claimed.